Long Horizon Coding Agent
It plans before it codes, remembers across sessions, and runs for hours at a fraction of the cost.
or install the CLI
Trusted by startups from
Trusted by startups from
Trusted by startups from
Plan, build, verify, ship. All of it, without you prompting it through every step.
1/5 Describe
Say what you want built.
2/5 Plan
Architecture and a roadmap, before any code.
3/5 Build
It writes the code and shows its work.
4/5 Verify
Nothing closes until it is proven.
5/5 Ship
It lands in your GitHub, with a live preview.
Connect it once, and every agent uses it while it builds.
Connections
One OAuth each. 25 services across 9 categories, and growing.
API keys
195 providers across 22 categories. Your billing, encrypted, injected at build time.
Agent skills
Standing instructions, injected into every conversation.
TypeScript everywhere
Always use TypeScript. No plain .js files, and no `any` without a comment saying why.
Our API conventions
Follow docs/api.md: plural resource paths, cursor pagination, errors as { code, message }.
Styling rules
Tailwind only, never CSS modules. Use the tokens in theme.css rather than raw hex values.
Every service is observable
Each service exposes /health and /ready, and logs structured JSON with a request id.
MCP servers
Any server, whole tool sets, over stdio, HTTP or SSE.
Every app pre.dev builds starts with all three working. No accounts to create, no keys to paste, no webhooks to register.
Hundreds of models from every major lab, from text and code to images and video.
Recorded in Restyle, which pre.dev built from one prompt. Try it live ↗
We outscore Claude Code and OpenCode on Terminal-Bench 2.0, a full model tier down, at a third of the cost.
Start for free. Upgrade anytime.
Alex Thompson
Founder & CEO at StartupFlow
It planned the whole build before touching code, then ran for hours and opened clean PRs. I reviewed diffs instead of babysitting prompts.
Maria Rodriguez
Staff Engineer at InnovateLab
It handled a multi-file refactor across our codebase without losing context halfway through. The persistent memory is the part other agents get wrong.
David Kim
CTO at TechVenture
Nothing gets marked done until it passes type checks, lint, and browser verification. We stopped getting "looks good but is broken" PRs.
Nicolas Zander
Founder at CleverCompliance
The upfront planning is what sold me. It decomposes the work into a dependency graph before it writes a line. Mindblowingly good.
Max Kryzhanovskiy
Lead Engineer at MosCreative
I've run it across three different model providers depending on the job. Being model-agnostic means I'm never locked in or overpaying on tokens.
Alex Thompson
Founder & CEO at StartupFlow
It planned the whole build before touching code, then ran for hours and opened clean PRs. I reviewed diffs instead of babysitting prompts.
Maria Rodriguez
Staff Engineer at InnovateLab
It handled a multi-file refactor across our codebase without losing context halfway through. The persistent memory is the part other agents get wrong.
David Kim
CTO at TechVenture
Nothing gets marked done until it passes type checks, lint, and browser verification. We stopped getting "looks good but is broken" PRs.
Nicolas Zander
Founder at CleverCompliance
The upfront planning is what sold me. It decomposes the work into a dependency graph before it writes a line. Mindblowingly good.
Max Kryzhanovskiy
Lead Engineer at MosCreative
I've run it across three different model providers depending on the job. Being model-agnostic means I'm never locked in or overpaying on tokens.
Alex Thompson
Founder & CEO at StartupFlow
It planned the whole build before touching code, then ran for hours and opened clean PRs. I reviewed diffs instead of babysitting prompts.
Maria Rodriguez
Staff Engineer at InnovateLab
It handled a multi-file refactor across our codebase without losing context halfway through. The persistent memory is the part other agents get wrong.
David Kim
CTO at TechVenture
Nothing gets marked done until it passes type checks, lint, and browser verification. We stopped getting "looks good but is broken" PRs.
Nicolas Zander
Founder at CleverCompliance
The upfront planning is what sold me. It decomposes the work into a dependency graph before it writes a line. Mindblowingly good.
Max Kryzhanovskiy
Lead Engineer at MosCreative
I've run it across three different model providers depending on the job. Being model-agnostic means I'm never locked in or overpaying on tokens.
Alex Thompson
Founder & CEO at StartupFlow
It planned the whole build before touching code, then ran for hours and opened clean PRs. I reviewed diffs instead of babysitting prompts.
Maria Rodriguez
Staff Engineer at InnovateLab
It handled a multi-file refactor across our codebase without losing context halfway through. The persistent memory is the part other agents get wrong.
David Kim
CTO at TechVenture
Nothing gets marked done until it passes type checks, lint, and browser verification. We stopped getting "looks good but is broken" PRs.
Nicolas Zander
Founder at CleverCompliance
The upfront planning is what sold me. It decomposes the work into a dependency graph before it writes a line. Mindblowingly good.
Max Kryzhanovskiy
Lead Engineer at MosCreative
I've run it across three different model providers depending on the job. Being model-agnostic means I'm never locked in or overpaying on tokens.
Cynthia Colmenares
Founder at Jack Frost Design
Every task lands as its own feature branch and PR. It never pushes to main, and it slots into the exact GitHub workflow our team already runs.
Ksenia Snegireva
CTO at Y-Tech
Each agent gets exactly the context it needs in an isolated sandbox, not our whole repo crammed into one window. The token savings on long runs are real.
Gleb Tertychnyi
Founder at WPWP
I pointed it at our existing repo and it reverse-engineered the architecture and kept building in our patterns. No starting from scratch.
Rick Wong
Founder at DataSaa
It kicked off the whole roadmap in one run and the agents worked through the dependency graph on their own. Genuinely long-horizon, not a glorified autocomplete.
Cynthia Colmenares
Founder at Jack Frost Design
Every task lands as its own feature branch and PR. It never pushes to main, and it slots into the exact GitHub workflow our team already runs.
Ksenia Snegireva
CTO at Y-Tech
Each agent gets exactly the context it needs in an isolated sandbox, not our whole repo crammed into one window. The token savings on long runs are real.
Gleb Tertychnyi
Founder at WPWP
I pointed it at our existing repo and it reverse-engineered the architecture and kept building in our patterns. No starting from scratch.
Rick Wong
Founder at DataSaa
It kicked off the whole roadmap in one run and the agents worked through the dependency graph on their own. Genuinely long-horizon, not a glorified autocomplete.
Cynthia Colmenares
Founder at Jack Frost Design
Every task lands as its own feature branch and PR. It never pushes to main, and it slots into the exact GitHub workflow our team already runs.
Ksenia Snegireva
CTO at Y-Tech
Each agent gets exactly the context it needs in an isolated sandbox, not our whole repo crammed into one window. The token savings on long runs are real.
Gleb Tertychnyi
Founder at WPWP
I pointed it at our existing repo and it reverse-engineered the architecture and kept building in our patterns. No starting from scratch.
Rick Wong
Founder at DataSaa
It kicked off the whole roadmap in one run and the agents worked through the dependency graph on their own. Genuinely long-horizon, not a glorified autocomplete.
Cynthia Colmenares
Founder at Jack Frost Design
Every task lands as its own feature branch and PR. It never pushes to main, and it slots into the exact GitHub workflow our team already runs.
Ksenia Snegireva
CTO at Y-Tech
Each agent gets exactly the context it needs in an isolated sandbox, not our whole repo crammed into one window. The token savings on long runs are real.
Gleb Tertychnyi
Founder at WPWP
I pointed it at our existing repo and it reverse-engineered the architecture and kept building in our patterns. No starting from scratch.
Rick Wong
Founder at DataSaa
It kicked off the whole roadmap in one run and the agents worked through the dependency graph on their own. Genuinely long-horizon, not a glorified autocomplete.
Post-training teams train coding models on hard, verified tasks built from private production code no model has seen. Each task ships with a runnable environment, a graded verifier, a hidden reference solution and solver trajectories.
AI labs license private codebases to train coding models. Connect a Git account, pick only the repos you're willing to license, and see what each one is worth.
What you'll earn
$6,000
per repo
non-exclusive licence, paid Net 15 after funds land
Example grades. Yours come from your own repos once they are analysed.
What pre.dev is, and the two things it does.
pre.dev is the Long Horizon Coding Agent. It plans before it codes, remembers across sessions, and runs autonomously for hours in the browser or your terminal, landing verified pull requests. It is model-agnostic. pre.dev also runs Browser Agents, the pre.dev AI Gateway, the Architect API and MCP server, and pre.dev Labs, which builds and sells RL coding tasks to frontier AI labs.
Yes. pre.dev Labs builds and sells RL Tasks to frontier AI labs for model post-training: hard, verified coding tasks built from private production code no model has seen. Each task ships with a runnable environment, a graded verifier, a hidden reference solution and solver trajectories. The pipeline produces 1,000+ new tasks every month from 10,000+ licensed private codebases. See pre.dev/rl-tasks and pre.dev/labs.
Post-training and RL teams at frontier AI labs that need long-horizon coding tasks with a verifiable reward, built from code that is not on public GitHub. Labs get tasks, not repository access, and can request samples or book a call with the founders before buying.
They are two offerings from the same company for different customers. The Long Horizon Coding Agent is for developers and teams who want software planned, built and verified for them. RL Tasks and Codebases are for AI labs training coding models. The agent is also the kind of long-horizon coding system the tasks are built to improve.
The Free plan is $0 with 20 credits, the coding agent in the browser and terminal, browser agents, and apps published for 7 days, with no credit card required. Paid plans start at Plus, $10 a month or $99 billed annually, with 100 monthly credits. Team and Enterprise plans are on the pricing page. RL Tasks are quoted per engagement, not part of these plans.
It plans it, builds it, proves it works, and ships it.
